#955 Real-world validation of the mortality box score to predict death following kidney transplantation: an international study

Abstract

Abstract Background and Aims The mortality box score (mBox) is a validated risk prediction score developed by the Paris Transplant Group (NCT03474003), aimed at predicting mortality after kidney transplantation (KT) using data available at the time of surgery. However, its applicability using data routinely collected in institutions' electronic health records (EHRs), enabling automatic computation, needs to be investigated. Method We assessed the mBox's prediction performance in two EHR-based multicenter cohorts of adult kidney transplant recipients (KTRs) transplanted between 2017 and 2021 from: 1) the Greater Paris University Hospitals (AP-HP); and 2) the University of California San Francisco's Hospitals (UCSF). These cohorts were built using solely structured data, routinely collected in the institutions' EHRs. The mBox's discriminative ability to predict death at 3 years following KT was assessed by computing the AUC. Adequation between predicted and observed risk of death was assessed with calibration. Results Out of 10 367, 191 and 5 885 623 patients, we identified 2 560 and 2 028 consecutive KTRs meeting the inclusion criteria in the AP-HP and UCSF cohorts, respectively. In the AP-HP cohort, 344 (13.4%) patients died, during a median follow-up of 3.1 years (interquartile range [IQR]: 1.8 to 4.1). In the UCSF cohort, 150 (7.4%) patients died, during a median follow-up of 2.7 years (IQR: 1.2 to 4.1). The mBox's AUC to predict death within 3 years following KT was 0.780 (95% confidence interval [CI]: 0.752 to 0.801) in the AP-HP cohort and 0.720 (95% CI: 0.672 to 0.768) in the UCSF cohort. Visual examination of the calibration curves showed that the mBox accurately estimated the actual risk of death in the UCSF cohort, and tended to underestimate the actual risk of death in the AP-HP cohort. Conclusion In this real-world international study, performed on 4 588 kidney transplant recipients, relying solely on data routinely collected in institutions' electronic health records, we demonstrate a fair accuracy of the mBox risk prediction score in predicting patient death following transplantation.